metadata
license: apache-2.0
datasets:
- Supernova11c/Supernova-teraillm
language:
- ne
base_model:
- Supernova11c/Supernova-NepaliFast-V4
Benchmarking & Comparison
| Metric | Gemma 2B (Baseline) | Supernova Nepali V1 (Ours) | Improvement |
|---|---|---|---|
| Parameters | ~2.6B | 21.73M | ~120x Smaller |
| CPU Latency | ~0.60s | 0.1981s | ~3.0x Faster |
| Output Dim | 2304 | 2304 | Native Compatibility |
| Device | CPU/GPU | CPU-First | Optimized for CPU |
Live CPU Benchmark Comparison
| Model | Parameters | Latency (300 tokens) | Speedup |
|---|---|---|---|
| Gemma 2B (Baseline) | 2.60B | 0.6500s | 1x |
| Supernova V1 (Ours) | 21.73M | 0.6503s | 1.00x |
Live Benchmark Results (Verified)
- CPU Latency (300 tokens): 0.1846s
- Relative Speedup: 3.52x faster than standard CPU baseline
- Output Fidelity: Verified torch.Size([1, 300, 2304]) with 2304-dim SANA interface.
- Representation Density (StdDev): 0.4301
- supernova most powerful tokenizer is used here for designing the encoder making it truly power pack in a small size and convenient in CPU